The global prevalence of diabetes is on the rise, emerging as a significant and pressing public health concern in the 21st century. Diabetes, characterized by either deficiency of insulin secretion, is a widespread chronic condition. According to recent statistics presented by the International Diabetes Federation (IDF), 8.8% of individuals affected by diabetes are approximately in the age range of 20–79 years. Notably, 46.5% of these cases go undetected, contributing to the alarming fact that ~5 million deaths annually are attributed to diabetes. Projections indicate that by 2040, the global diabetic population will reach 642 million (Ershadi and Seifi, 2020).

In the past few years, the exploration of medical data collected through data mining and machine learning techniques has captured the attention of numerous researchers. Researchers have worked with the Bayesian Belief Network due to its nature and suitability in the clinical world (Kharya et al., 2022). In the clinical sector, the Bayesian network projects conditional dependencies and independencies between the various symptoms of illness (Fazel Zarandi et al., 2018; Braik et al., 2023a). For example, “physical inactivity” may cause “obesity,” which may raise “blood glucose.” Moreover, BBN plays a significant role in understanding the association between symptoms. In this study, to improve the accuracy, BBN is reconstructed to incorporate one characteristic of the clinical world, i.e., “not all symptoms are equally important for prediction.” Every symptom has different predicting capabilities for any illness in the clinical world, for example, considering the symptoms of heart diseases such as chest pain, fainting, fatigue, shortness of breath, and swollen feet. Here, each symptom has a different impact on the severity of the disease. Therefore, it cannot be given equal weightage. According to the above example, chest pain impact is the highest among all the symptoms. Moreover, an innovative idea of assigning different weights to different symptoms based on the predictive capability is incorporated into this study. This study proposes a new concept, formula, and pseudocode to create a new Weighted Bayesian Belief Network (WBBN) model using a clinical dataset. In this model, the weighted concept is introduced and implemented with BBN.

The suitability of the BBN in the clinical domain as the best predictive computational model is presented in the study mentioned in the reference (Liu et al., 2018; Shen et al., 2018; Simsek et al., 2020; Braik et al., 2023b). After studying the impact of different disease symptoms, the major work is to find the correlation or association between the symptoms. The importance of the ARM method in discovering a correlation or association between attributes/symptoms is to predict the class label in the form of Class Association Rules (CARs). In conventional rule mining, the significance of a rule is typically determined by the count of item sets within a database. Traditional mining rules rely on support and confidence measures to identify frequent item sets, assuming that all items have equal significance (Topuz et al., 2018).

In contrast, social science, medical, and business market researchers hold distinct perspectives. A rule's significance is contingent on quantitative aspects, such as the frequency of an item in a database, and qualitative elements, involving human interpretation rather than solely relying on database metrics. Weights can be employed to depict the impact of symptoms within a dataset (Kumar et al., 2019).

This study introduces a novel classifier, WBBN, which employs Weighted Bayesian_class Association Rules (WBAR) to construct the computational model utilizing clinical datasets. This methodology initially allocates weights to various symptoms or attributes based on their predictive capacities. Next, the main focus is given to the WARM technique, which discovers the relationship between Weighted Two attributes and Weighted Multi attributes to represent hidden patterns and new knowledge. Then, the Weighted_class Association Rules (WAR) are extracted. The consequence of a rule is the 'class label' using interesting measures such as Weighted Support and Weighted Confidence on setting minimum threshold values. Bayesian theory is applied to the Weighted_class Association Rules to produce WBAR, utilizing Weighted Bayesian Confidence (WBC) and Weighted Bayesian Lift (WBL), which are employed to construct the WBBN. Subsequently, experiments were conducted on established clinical datasets to assess the accuracy of the performance of WBBN.

A brief literature survey is conducted on the research performed in the clinical sector of the last 5 years on different classifiers using the PIDD UCI machine learning dataset, as shown in Table 1. The classifiers considered are Naïve Bayes, Neural Network, Support Vector Machine, and Decision Tree. Accuracies of all the work are shown in Table 1. The literature survey shows that the Bayesian Belief Network classifier is a promising area to work on, and it shows better results when applied to the PIDD UCI machine learning dataset.

Experiments are performed on clinical datasets such as diabetes using the Pima Indian Diabetics Dataset (PIDD) from the UCI machine learning data repository (Pima Indian Diabetes Dataset UCI.-ML Repository, 2023). The distribution and details of data are shown in Table 2.

The clinical datasets are acquired from the standard UCI archive. The discretized version is obtained through Liverpool University Computer Science-Knowledge Discovery in Data (LUCS-KDD). Discretization/Normalized (DN) software is utilized to convert data files from the UCI archive, which is an appropriate format for ARM applications. In this context, discretization refers to converting numeric attributes into categorical ones.

The procedural approach of the WBBN model proposed in this research is systematically elucidated through a step-by-step representation depicted in the workflow diagram, as shown in Figure 1. This diagram outlines the sequential processes employed in investigating the proposed study.

Following extracting the clinical dataset using the UCI archive and obtaining its discretized form, the subsequent step involves calculating the weights of the {attribute, value} pair using an automated method. Weights can be calculated using domain knowledge-based weight calculation, automated weight calculation, and semi-automated weight calculation (Alwidian et al., 2018). This study calculates weights using the rank-based weight assignment method (Kharya and Soni, 2016). The choice of the rank-based weight assignment method for computing weights of {attribute, value} pairs in the WBBN model can contribute to its robustness and interpretability such as ranking importance, simplicity, and transparency, especially in clinical decision support systems. This ranking clearly indicates which features are most influential in making predictions, enhancing the interpretability of the model. This attribute ranking facilitates clinicians' understanding of the model's decision-making process in clinical decision support systems, where transparency and understandability are crucial.

Different ranking methods may yield varying results in terms of attribute importance. The choice of ranking algorithm or criteria can impact the weights assigned to attribute-value pairs and, consequently, the performance of the model. Careful consideration and validation of the chosen ranking method are necessary to ensure its suitability for the specific clinical domain and dataset.

The whole procedure for building the WBBN model is explained further using definitions, formulas, and pseudocode. The end product is a clinical decision support system used as a predictive model termed the WBBN model.

A benchmark medical dataset related to PIDD is utilized to assess the effectiveness of the WBBN model by applying weighted Bayesian_class association rules. The construction of the model involves utilizing Java version 1.8 for the front end and MySql 8 as the backend tool. Different proportions of the dataset are employed to train and test to learn the innovative predictive model, employing various thresholds for Min_Weighted support and Min_Weighted confidence. The outcomes of four distinct experimental scenarios are tabulated using Tables 3–6. In this study, the WBBN model undergoes thorough training and testing using a distinct distribution of the PIDD dataset comprising 768 records. The primary evaluative parameter employed in this study is accuracy, which is the correctness of predictions made by a predictive model for diabetes diagnosis or classification (Nayak et al., 2023; Panigrahi et al., 2023; Pati et al., 2023).

The WBBN model using the clinical dataset is evaluated regarding the number of strong rules and accuracy. Table 7 shows the results of the WBBN model using three clinical datasets: the Pima Indian Diabetic dataset, the Heart Disease dataset, and the Breast Cancer dataset.WBBN model, built using seven strong rules of PIDD, achieves an accuracy of 95.8% when the model is trained using 80% of the dataset, and tested on 20% with provided Min_Threshold. Again, the WBBN model built using seven strict rules of the heart disease dataset acquires an accuracy of 92.7% when the model is trained on 70% of the dataset and tested on 30% of the data with the provided Min_Threshold. Similarly, the WBBN model achieves the highest accuracy of 97.18%, as shown in Table 7. The graphical representation of the results is shown in Figure 3.

Now, the exhaustive comparison of the proposed model WBBN is done with existing classifiers developed in recent years using the same diabetic dataset. The comparative study shows that the proposed model WBBN gives outstanding results, as shown in Table 8. In this comparison, the WBBN model achieves the highest accuracy of 95.8%. This could be attributed to its ability to capture complex relationships between features in the dataset. The superior performance of WBBN indicates the potential effectiveness of probabilistic modeling for classification tasks.

Here, Naive Bayes achieves an accuracy of 80%, which is notably lower than WBBN. The simplicity of Naive Bayes, relying on the assumption of feature independence, may not fully capture the complex relationships in the dataset, leading to lower accuracy.

Again, SVM performs admirably with an accuracy of 94.5%, slightly below WBBN. SVM's ability to identify complex decision boundaries in high-dimensional spaces might contribute to its competitive performance. MLP and ANN achieve accuracies of 86.08% and 91%, respectively.

While these neural network models demonstrate reasonable performance, they fall short of WBBN, possibly due to suboptimal architecture or training parameters. Decision Tree achieves an accuracy of 82.14%, which is relatively lower than other models. Inherent limitations of Decision Trees in capturing complex relationships and tendency to overfit might contribute to their lower accuracy.

To increase the visualization of a comparison, a graph has been plotted for PIDD clinical datasets on various existing classifiers such as Naïve Bayes, SVM, MLP-NN, K-NN, Random forest, and Decision Tree, which are commonly used in the clinical industry, as shown in Figure 4. In conclusion, while WBBN demonstrates superiority in accuracy over other classifiers on the Pima Indian Diabetic Dataset, further research is needed to explore its interpretability, scalability, and generalization capabilities. Additionally, addressing the limitations and challenges encountered during the evaluation processes, such as dataset bias and class imbalance, would provide valuable insights for enhancing the effectiveness of classifiers in practical applications, particularly in healthcare and related domains.

The construction of the WBBN model involves incorporating the property of the clinical dataset, specifically the notion that “not all symptoms are equally important for prediction.” This is achieved by assigning varying weights to attributes based on their predictive capabilities. The experimental findings demonstrate that the weighted concept contributes to achieving higher accuracy in the clinical domain compared with other available predictive models. The experiments were conducted using three distinct clinical datasets, namely, the Breast Cancer dataset, Heart Disease dataset, and PIDD from the UCI archive, yielding outstanding results. This achievement represents a significant contribution to the medical sector. In the future, the innovative weighted model could be extended to non-clinical datasets to benefit other sectors. Extending the weighted model to non-clinical datasets requires careful consideration of domain-specific characteristics, data quality, feature engineering, interpretability, generalization, ethical considerations, and scalability. While the weighted model may offer advantages in terms of performance and interpretability, addressing the challenges and limitations inherent in applying the model to diverse datasets beyond the medical domain is essential for its successful adoption and deployment in non-clinical sectors. Additionally, addressing the “Sharp Boundary problem in the medical field” could be achieved by incorporating fuzzy theory and developing a fuzzy weighted model.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.

SK: Writing—original draft, Writing—review & editing. SS: Writing—original draft, Writing—review & editing. APat: Conceptualization, Investigation, Writing—review & editing, Writing—original draft. APan: Conceptualization, Investigation, Writing—review & editing, Writing—original draft. JG: Conceptualization, Data curation, Investigation, Methodology, Visualization, Writing—review & editing, Writing—original draft. HQ: Conceptualization, Data curation, Investigation, Methodology, Software, Writing—original draft, Writing—review & editing. SM: Conceptualization, Investigation, Software, Writing—review & editing, Writing—original draft. DN: Data curation, Investigation, Writing—review & editing, Writing—original draft. TS: Conceptualization, Data curation, Writing—review & editing, Writing—original draft.